Key Points
- A $1 billion Series F first close values SambaNova at $11 billion.
- General Atlantic led the round, five months after a $350 million Series E.
- JPMorgan Chase named SambaNova an inference-infrastructure partner for on-prem AI.
- SoftBank is the first deployment partner for the SN50 chip, shipping in H2 2026.
SambaNova has raised $1 billion at an $11 billion valuation, and the more telling number is five — as in the number of months since its last round. The Palo Alto chipmaker closed a $350 million Series E in February. The market has repriced it in less than half a year, and the reason is where the AI bill actually lands.
The Series F first close was led by General Atlantic, with T. Rowe Price, Capital Group, BlackRock, Vista Equity Partners, Seligman Ventures, and the Qatar Investment Authority also in. SambaNova, founded in 2017 and led by CEO Rodrigo Liang, builds reconfigurable dataflow chips tuned for inference — the SN40L that shipped in 2023 and the next-generation SN50 it unveiled in February, which packs an entire rack's worth of memory around a single design so it can run trillion-parameter models fast.
SambaNova's was not the only nine-figure round of the week: Neko Health raised $700 million led by Lightspeed at a valuation near $7 billion for preventive body scans — the same investor appetite, aimed at the least software-like business imaginable.
The customers doing the talking
The raise arrived with two names that matter more than the investor list. JPMorgan Chase selected SambaNova as an "inference-infrastructure partner," deploying its SN40L and SN50 systems to run AI inference on-premises inside the bank rather than in a public cloud. And SoftBank signed on as the first deployment partner for the SN50, with shipments expected in the second half of 2026.
The read
Training a model is a one-time capital cost; inference is the bill that arrives every time the model answers. That distinction is the entire investment case here, and it explains the speed of the markup. The money is chasing the recurring layer — the compute that gets paid on every query — and betting a share of it can be pulled out of Nvidia's hands. JPMorgan's choice is the sharpest tell: a regulated bank that wants inference it can run on its own floor, on silicon that isn't a GPU, is exactly the customer SambaNova needs to prove the thesis. It fits a 2026 pattern of capital flooding the layer beneath the models, from Baseten's $1.5 billion Series F for inference infrastructure to Qualcomm's talks to buy the chip startup Tenstorrent, all of it circling the same question of who supplies compute once Nvidia is no longer the only option. Even Nvidia is pushing into new markets to stay ahead of exactly this pressure. The same logic priced Etched, whose $700 million round at a $21 billion valuation was led by Jane Street after the trading firm bought and deployed one of its first racks.
Source: TechCrunch.
Comments